Muhammad Azam
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Muhammad Azam has a total of 27 co-authors. Here are the top co-authors by number of shared publications:
- Raheem Sarwar
OTEHM, Manchester Metropolitan University, Manchester, UK
4 shared publications - Nadeem Ahmad
Faculty of CS&IT, Superior University, Lahore, Pakistan
3 shared publications - Fahad Sabah
College of Computer Science, Beijing University of Technology, Beijing, China
2 shared publications - Abdul Manan
Faculty of CS&IT, Superior University, Lahore, Pakistan
1 shared publication - Abdul Qadir Khan
College of Computer Science, Beijing University of Technology, Beijing, China
1 shared publication - Abdul Rasheed
Department of Molecular Neuroscience University of Toyama, Japan
1 shared publication
Latest Publications
- 24 July 2026
- Published by Springer Nature in SN Computer Science
- Vol. 7 (6) , 650
Mapping the Intellectual Landscape of Personalized Federated Learning: A Bibliometric Analysis
- 24 July 2026
- Published by Springer Nature in SN Computer Science
- Vol. 7 (6) , 650
Mapping the Intellectual Landscape of Personalized Federated Learning: A Bibliometric Analysis
- 20 July 2026
- Published by Springer Nature in Journal of Robotic Surgery
- Vol. 20 (1) , 699
Explainable machine learning for automated robotic surgical skill assessment using rich kinematic features
Robotic-assisted surgery (RAS) extends minimally invasive surgery by restoring dexterity, tremor filtration, and ergonomic console control compared with open procedures and conventional laparoscopy. Objective, scalable skill assessment from console kinematics is therefore clinically relevant for competency-based training. Existing JIGSAWS studies often rely on random trial splits, isolated handcrafted features, or black-box models without subject-independent validation or transparent explanations. This study presents an explainable machine learning framework for automated skill classification from kinematic data alone. Using the public JHU-ISI Gesture and Skill Assessment Working Set (JIGSAWS), comprising 103 trials from 40 subjects across three dry-lab tasks, 633 trial-level kinematic features were extracted and six supervised classifiers were compared under leave-one-subject-out cross-validation (LOSO-CV). Random Forest (RF) with feature standardization was selected as the primary model; interpretability was assessed with SHapley Additive exPlanations (SHAP). Key gaps addressed include: subject-independent evaluation on all trials, systematic classifier comparison on identical features, collinearity-aware feature analysis, global and local SHAP attribution, and ablations (mean/standard-deviation features, segment-level modeling, PCA reduction, deep sequences, cross-task transfer). Under pooled LOSO-CV, RF with advanced kinematic features achieved 97.1% accuracy and Cohen’s \(\kappa =0.955\) (macro-F1 \(=0.971\); balanced accuracy \(=96.4\%\)), outperforming PCA-reduced features (92.2%) and segment-level modeling (88.3%). SHAP analysis identified jerk and velocity variability in master-tool channels as dominant discriminators. Cross-task evaluation showed strong in-task performance (97–100%) but reduced off-task transfer (61–95%). These benchmark findings may inform explainable kinematics-based training analytics; external multi-center validation remains required before clinical deployment.- 20 July 2026
- Published by Springer Nature in Journal of Robotic Surgery
- Vol. 20 (1) , 699
Explainable machine learning for automated robotic surgical skill assessment using rich kinematic features
Robotic-assisted surgery (RAS) extends minimally invasive surgery by restoring dexterity, tremor filtration, and ergonomic console control compared with open procedures and conventional laparoscopy. Objective, scalable skill assessment from console kinematics is therefore clinically relevant for competency-based training. Existing JIGSAWS studies often rely on random trial splits, isolated handcrafted features, or black-box models without subject-independent validation or transparent explanations. This study presents an explainable machine learning framework for automated skill classification from kinematic data alone. Using the public JHU-ISI Gesture and Skill Assessment Working Set (JIGSAWS), comprising 103 trials from 40 subjects across three dry-lab tasks, 633 trial-level kinematic features were extracted and six supervised classifiers were compared under leave-one-subject-out cross-validation (LOSO-CV). Random Forest (RF) with feature standardization was selected as the primary model; interpretability was assessed with SHapley Additive exPlanations (SHAP). Key gaps addressed include: subject-independent evaluation on all trials, systematic classifier comparison on identical features, collinearity-aware feature analysis, global and local SHAP attribution, and ablations (mean/standard-deviation features, segment-level modeling, PCA reduction, deep sequences, cross-task transfer). Under pooled LOSO-CV, RF with advanced kinematic features achieved 97.1% accuracy and Cohen’s \(\kappa =0.955\) (macro-F1 \(=0.971\); balanced accuracy \(=96.4\%\)), outperforming PCA-reduced features (92.2%) and segment-level modeling (88.3%). SHAP analysis identified jerk and velocity variability in master-tool channels as dominant discriminators. Cross-task evaluation showed strong in-task performance (97–100%) but reduced off-task transfer (61–95%). These benchmark findings may inform explainable kinematics-based training analytics; external multi-center validation remains required before clinical deployment.- Open Access
- 11 June 2026
- Published by Springer Nature in Journal of Robotic Surgery
- Vol. 20 (1) , 583
Correction: Technical advances in robotic retinal surgery: a systematic review and future research directions
- 11 June 2026
- Published by Springer Nature in Journal of Robotic Surgery
- Vol. 20 (1) , 583
Open AccessCorrection: Technical advances in robotic retinal surgery: a systematic review and future research directions
- Open Access
- 11 June 2025
- Published by Slovenian Association Informatika in Informatica
- Vol. 49 (23)
Uncertainty-Aware Energy Consumption Forecasting Using LSTM Networks with Monte Carlo Dropout
Accurate forecasting of energy consumption is critical for effective resource management and sustainability in the energy sector. This paper presents an uncertainty-aware deep learning approach using Long Short-Term Memory (LSTM) networks with Monte Carlo Dropout to enhance prediction accuracy and quantify uncertainty. Our model is trained on hourly energy consumption data from the PJM Electricity Market (2015–2020), preprocessed via temporal feature engineering (hour-of-day, day-of-week, month), linear interpolation for missing values, and Z-score-based outlier removal. The proposed framework achieves RMSE: 5005.93, MAE: 4063.75, and MAPE: 13% on the test set, outperforming benchmark models like ARIMA (RMSE: 6500) and Exponential Smoothing (RMSE: 7200). By integrating Monte Carlo Dropout during inference, we generate probabilistic forecasts with 95% confidence intervals, enabling stakeholders to assess prediction reliability. Cross-validation results (average RMSE: 16015.68) highlight the model’s robustness to temporal variability. Our work demonstrates that LSTM networks with uncertainty quantification significantly improve energy forecasting accuracy, offering actionable insights for grid management and policy decisions.
- 11 June 2025
- Published by Slovenian Association Informatika in Informatica
- Vol. 49 (23)
Open AccessUncertainty-Aware Energy Consumption Forecasting Using LSTM Networks with Monte Carlo Dropout
Accurate forecasting of energy consumption is critical for effective resource management and sustainability in the energy sector. This paper presents an uncertainty-aware deep learning approach using Long Short-Term Memory (LSTM) networks with Monte Carlo Dropout to enhance prediction accuracy and quantify uncertainty. Our model is trained on hourly energy consumption data from the PJM Electricity Market (2015–2020), preprocessed via temporal feature engineering (hour-of-day, day-of-week, month), linear interpolation for missing values, and Z-score-based outlier removal. The proposed framework achieves RMSE: 5005.93, MAE: 4063.75, and MAPE: 13% on the test set, outperforming benchmark models like ARIMA (RMSE: 6500) and Exponential Smoothing (RMSE: 7200). By integrating Monte Carlo Dropout during inference, we generate probabilistic forecasts with 95% confidence intervals, enabling stakeholders to assess prediction reliability. Cross-validation results (average RMSE: 16015.68) highlight the model’s robustness to temporal variability. Our work demonstrates that LSTM networks with uncertainty quantification significantly improve energy forecasting accuracy, offering actionable insights for grid management and policy decisions.
- Open Access
- 11 April 2025
- Published by Superior University, Lahore in Journal of Innovative Computing and Emerging Technologies
- Vol. 5 (1)
Applying Multiple Feature selection strategies to Boost the Effectiveness of Machine Learning and Deep Learning Models for Angiography Outcome Predictions
Angiography serves as a vital imaging technique for diagnosing and treating cardiovascular disease by enabling precise visualization of coronary arteries. This study investigates the impact of multiple feature selection strategies on enhancing the predictive accuracy of machine learning (ML) and deep learning (DL) models in classifying angiography outcome. The dataset consists of 976 angiography videos, categorized into three classes: Normal, Coronary Artery Bypass Grafting (CABG), and Angioplasty. [1] To ensure data reliability, cases related to percutaneous Trans venous Mitral Commissurotomy (PTMC) and insufficient patient records were excluded. A Comparative analysis was conducted using traditional ML models such as Support Vector Machines (SVM), Random Forest (RF), and Gradient Boosting Machines (GBM) alongside deep learning architecture like Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks. [2] Various feature selection techniques, including Principle Component Analysis (PCA), Recursive Feature Elimination (RFE), and Mutual Information (MI), were employed to enhance model interpretability and efficiency. [3] Results demonstrate that while ML models achieved moderate accuracy ranging between 85% and 90% and DL models significantly outperformed them, reaching an accuracy of 91%. The superior performance of DL approaches can be attributed to their ability to automatically extract hierarchical spatial and temporal features from angiography videos, whereas ML models rely heavily on handcrafted features, limiting their effectiveness [4]. These findings emphasize the critical role of feature selection and deep learning methodologies in improving angiography outcome predictions, ultimately contributing to Al-driven advancements in cardiovascular diagnostics. The study underscores the potential of DL-based automated angiography assessment in supporting clinical decision-making, leading to improved patient management and early intervention strategies. [5]
- 11 April 2025
- Published by Superior University, Lahore in Journal of Innovative Computing and Emerging Technologies
- Vol. 5 (1)
Open AccessApplying Multiple Feature selection strategies to Boost the Effectiveness of Machine Learning and Deep Learning Models for Angiography Outcome Predictions
Angiography serves as a vital imaging technique for diagnosing and treating cardiovascular disease by enabling precise visualization of coronary arteries. This study investigates the impact of multiple feature selection strategies on enhancing the predictive accuracy of machine learning (ML) and deep learning (DL) models in classifying angiography outcome. The dataset consists of 976 angiography videos, categorized into three classes: Normal, Coronary Artery Bypass Grafting (CABG), and Angioplasty. [1] To ensure data reliability, cases related to percutaneous Trans venous Mitral Commissurotomy (PTMC) and insufficient patient records were excluded. A Comparative analysis was conducted using traditional ML models such as Support Vector Machines (SVM), Random Forest (RF), and Gradient Boosting Machines (GBM) alongside deep learning architecture like Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks. [2] Various feature selection techniques, including Principle Component Analysis (PCA), Recursive Feature Elimination (RFE), and Mutual Information (MI), were employed to enhance model interpretability and efficiency. [3] Results demonstrate that while ML models achieved moderate accuracy ranging between 85% and 90% and DL models significantly outperformed them, reaching an accuracy of 91%. The superior performance of DL approaches can be attributed to their ability to automatically extract hierarchical spatial and temporal features from angiography videos, whereas ML models rely heavily on handcrafted features, limiting their effectiveness [4]. These findings emphasize the critical role of feature selection and deep learning methodologies in improving angiography outcome predictions, ultimately contributing to Al-driven advancements in cardiovascular diagnostics. The study underscores the potential of DL-based automated angiography assessment in supporting clinical decision-making, leading to improved patient management and early intervention strategies. [5]
- Open Access
- 26 December 2024
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 81 (1) , 1-35
Deep deterministic policy gradients with a self-adaptive reward mechanism for image retrieval
Traditional image retrieval methods often face challenges in adapting to varying user preferences and dynamic datasets. To address these limitations, this research introduces a novel image retrieval framework utilizing deep deterministic policy gradients (DDPG) augmented with a self-adaptive reward mechanism (SARM). The DDPG-SARM framework dynamically adjusts rewards based on user feedback and retrieval context, enhancing the learning efficiency and retrieval accuracy of the agent. Key innovations include dynamic reward adjustment based on user feedback, context-aware reward structuring that considers the specific characteristics of each retrieval task, and an adaptive learning rate strategy to ensure robust and efficient model convergence. Extensive experimentation with the three distinct datasets demonstrates that the proposed framework significantly outperforms traditional methods, achieving the highest retrieval accuracy having 3.38%, 5.26%, and 0.21% improvement overall as compared to the mainstream models over DermaMNIST, PneumoniaMNIST, and OrganMNIST datasets, respectively. The findings contribute to the advancement of reinforcement learning applications in image retrieval, providing a user-centric solution adaptable to various dynamic environments. The proposed method also offers a promising direction for future developments in intelligent image retrieval systems.
- 26 December 2024
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 81 (1) , 1-35
Open AccessDeep deterministic policy gradients with a self-adaptive reward mechanism for image retrieval
Traditional image retrieval methods often face challenges in adapting to varying user preferences and dynamic datasets. To address these limitations, this research introduces a novel image retrieval framework utilizing deep deterministic policy gradients (DDPG) augmented with a self-adaptive reward mechanism (SARM). The DDPG-SARM framework dynamically adjusts rewards based on user feedback and retrieval context, enhancing the learning efficiency and retrieval accuracy of the agent. Key innovations include dynamic reward adjustment based on user feedback, context-aware reward structuring that considers the specific characteristics of each retrieval task, and an adaptive learning rate strategy to ensure robust and efficient model convergence. Extensive experimentation with the three distinct datasets demonstrates that the proposed framework significantly outperforms traditional methods, achieving the highest retrieval accuracy having 3.38%, 5.26%, and 0.21% improvement overall as compared to the mainstream models over DermaMNIST, PneumoniaMNIST, and OrganMNIST datasets, respectively. The findings contribute to the advancement of reinforcement learning applications in image retrieval, providing a user-centric solution adaptable to various dynamic environments. The proposed method also offers a promising direction for future developments in intelligent image retrieval systems.
- Open Access
- 3 October 2024
- Published by Superior University, Lahore in Journal of Innovative Computing and Emerging Technologies
- Vol. 4 (2)
A Study on the Effectiveness of Fine-Tuning VGG16 Versus Classical Machine Learning and Deep Learning Models for Skin Burn Detection
Accurate skin burn classification is essential for successful medical diagnosis and treatment. This study focuses on the skin burn dataset of 1000 training images and 150 validation images where recognition performance of various classical machine learning algorithms as well as deep learning models has been analyzed. The classical models of deep learning that have been covered in this paper include, namely, Support Vector Machine (SVM), Random Forest (RF), K- Nearest Neighbors (KNN), and Gradient Boosting Machines (GBM). Also , we evaluate the performance of Convolutional Neural Networks (CNNs) and advanced transfer learning techniques using VGG16 for image recognition, classification and segmentation. The result appear convincing although classical machine learning methods offered valuable insights, their performance was undermined by deep learning approaches. In particular, the basic CNN model demonstrated average accuracy, but it was the fine-tuning of the VGG16 model that achieved the highest validation accuracy of 82.67%. The fine-tuned VGG16 model significantly outperformed classical models which is easy to fit large dataset for image classification. The fine-tuned VGG16 model which had accuracy scores increase from 65.33% to 73.33%, and hybrid models showed different degrees of improvement. The study also demonstrates how fine-tuned deep learning models are superior to some approaches in machine learning and offers detailed comparisons of various classification methodologies. The foundations of this study are laid down with the aim of increasing the efficacy of the automatic diagnosis of skin burn analysis and improving the efficiency of medical image analysis.
- 3 October 2024
- Published by Superior University, Lahore in Journal of Innovative Computing and Emerging Technologies
- Vol. 4 (2)
Open AccessA Study on the Effectiveness of Fine-Tuning VGG16 Versus Classical Machine Learning and Deep Learning Models for Skin Burn Detection
Accurate skin burn classification is essential for successful medical diagnosis and treatment. This study focuses on the skin burn dataset of 1000 training images and 150 validation images where recognition performance of various classical machine learning algorithms as well as deep learning models has been analyzed. The classical models of deep learning that have been covered in this paper include, namely, Support Vector Machine (SVM), Random Forest (RF), K- Nearest Neighbors (KNN), and Gradient Boosting Machines (GBM). Also , we evaluate the performance of Convolutional Neural Networks (CNNs) and advanced transfer learning techniques using VGG16 for image recognition, classification and segmentation. The result appear convincing although classical machine learning methods offered valuable insights, their performance was undermined by deep learning approaches. In particular, the basic CNN model demonstrated average accuracy, but it was the fine-tuning of the VGG16 model that achieved the highest validation accuracy of 82.67%. The fine-tuned VGG16 model significantly outperformed classical models which is easy to fit large dataset for image classification. The fine-tuned VGG16 model which had accuracy scores increase from 65.33% to 73.33%, and hybrid models showed different degrees of improvement. The study also demonstrates how fine-tuned deep learning models are superior to some approaches in machine learning and offers detailed comparisons of various classification methodologies. The foundations of this study are laid down with the aim of increasing the efficacy of the automatic diagnosis of skin burn analysis and improving the efficiency of medical image analysis.